You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Problem — The framed app, the connections bridge, and the vendored binary pin still point at the fork lineage: the panel iframes the ENGINE origin (which only works while the vendored fork binary serves the amicode app at its root), the overlay's /amicode/* fetches ride the server origin, and the vendored pin still names the fork's v1.18.10 release instead of stock v1.18.29. The fork cannot retire until the panel frames the extension service and the pin flips. Approach — Execute the M3 cutover the service was built for: every consumer of the engine-origin UI flips to the service origin (chat panel, deck panes, connections bridge), the service's parallel-run language retires, and the vendored binary pin flips to the stock canonical v1.18.29 release assets the overlay was re-based onto. Scope — in: the consumer flip (panel + deck + bridge call sites); the vendored pin flip with any fixture/buildinfo re-records it demands; retiring the parallel-run comments; the end-to-end proof against stock. · out: the shelf/proxy itself (#822, prerequisite); the upstream Gemini fix (Aaron's call, separate); any new surface.
Acceptance Criteria
The chat panel and every deck pane iframe the service origin (not the engine origin), bootstrapping with a credential the service accepts
The overlay app works end-to-end framed from the service origin against the stock v1.18.29 engine: engine API + SSE through the proxy, /amicode/* surfaces (home cards, widgets, solver-mode toggle) from the service routes
The connections bridge targets the service (its connections routes are native there) and the vendored terminal's service URL export remains the probe path
The vendored binary pin names the stock canonical v1.18.29 release assets (sha-verified per the fetch path's discipline), and every pin-dependent record (contract fixtures, buildinfo, golden files) is re-recorded from it
The parallel-run wiring language is retired honestly (the M3 cutover note lands where the "consumers stay on the fork" comments live)
The full local gate is green: typecheck AND the extension suite (the known environmental set excluded is documented, not assumed)
Testing Decisions
The service harness from #822 carries the framed-path contract tests; the pin flip re-records through the existing fetch/fixture machinery (the re-base slice's re-record precedent, PR #797). The boot proof from #822 re-runs pointed at the stock engine — the flip's end-to-end evidence.
Key Decisions
The pin flip is part of THIS slice (the flip without it is a no-op against the fork binary) but is the riskiest part: if fixture re-records reveal an unplanned base dependence, split it into its own follow-up PR rather than forcing it — honest sequencing over a bundled red.
The solver-mode route parity gap was already closed (the route port + golden re-record landed); the flip has no known parity gaps left.
Constraints & Invariants
Never merge non-green; the fork retires ONLY after this lands and CI proves the stock path — no early fork-side deletions.
#822 (the shelf — prerequisite); the parallel-run service slice (#451 M1) and its wiring comments; the re-base leg (PR #797 + #801/#802/#803) that landed the overlay on canonical and named these two blockers; the phase-2 disposition (consumers flip "when the service URL is framed").
Source
Fork-retirement blocker #2 from the harness-agnostic campaign's loop 28d. Part of the fleet-amicissimo-forksplit synthesis campaign. Blocked by #822.
Notes
Label hitl — human merge gate. Worktree off origin/main; the shared checkout's branch is never touched.
Important
Problem — The framed app, the connections bridge, and the vendored binary pin still point at the fork lineage: the panel iframes the ENGINE origin (which only works while the vendored fork binary serves the amicode app at its root), the overlay's /amicode/* fetches ride the server origin, and the vendored pin still names the fork's v1.18.10 release instead of stock v1.18.29. The fork cannot retire until the panel frames the extension service and the pin flips.
Approach — Execute the M3 cutover the service was built for: every consumer of the engine-origin UI flips to the service origin (chat panel, deck panes, connections bridge), the service's parallel-run language retires, and the vendored binary pin flips to the stock canonical v1.18.29 release assets the overlay was re-based onto.
Scope — in: the consumer flip (panel + deck + bridge call sites); the vendored pin flip with any fixture/buildinfo re-records it demands; retiring the parallel-run comments; the end-to-end proof against stock. · out: the shelf/proxy itself (#822, prerequisite); the upstream Gemini fix (Aaron's call, separate); any new surface.
Acceptance Criteria
Testing Decisions
The service harness from #822 carries the framed-path contract tests; the pin flip re-records through the existing fetch/fixture machinery (the re-base slice's re-record precedent, PR #797). The boot proof from #822 re-runs pointed at the stock engine — the flip's end-to-end evidence.
Key Decisions
Constraints & Invariants
Prior Art
#822 (the shelf — prerequisite); the parallel-run service slice (#451 M1) and its wiring comments; the re-base leg (PR #797 + #801/#802/#803) that landed the overlay on canonical and named these two blockers; the phase-2 disposition (consumers flip "when the service URL is framed").
Source
Fork-retirement blocker #2 from the harness-agnostic campaign's loop 28d. Part of the fleet-amicissimo-forksplit synthesis campaign. Blocked by #822.
Notes
Label hitl — human merge gate. Worktree off origin/main; the shared checkout's branch is never touched.